Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering

Jonas Kulhanek, Marie-Julie Rakotosaona, Fabian Manhardt, Christina Tsalicoglou, Michael Niemeyer, Torsten Sattler, Songyou Peng, Federico Tombari

TL;DR

This work tackles real-time rendering of large-scale 3D Gaussian Splatting on memory-constrained devices by introducing LODGE, a level-of-detail framework that builds multiple Gaussian sets $\mathcal{G}^{(l)}$ and precomputes per-chunk active Gaussians to avoid per-frame recomputation. It combines depth-aware 3D smoothing and importance pruning to generate coarser levels, uses a defined active set $\tilde{\mathcal{G}}(\mathbf{c}) = \bigcup_{l=0}^{L-1} \{ g_i^{(l)} : d_l \le \|\mu_i^{(l)}-\mathbf{c}\|_2 < d_{l+1} \}$, and partitions the scene into chunks via K-means over training camera positions; opacity blending between the two nearest chunks ensures smooth transitions. The approach also employs visibility filtering and a two-chunk blending scheme to preserve visual fidelity while dramatically reducing memory and computation. Evaluations on outdoor Hierarchical 3DGS and indoor Zip-NeRF datasets show a favorable quality-speed trade-off, enabling real-time mobile rendering with a substantially smaller memory footprint compared to prior methods. These advances make large-scale 3D reconstructions more accessible for AR/VR and navigation tasks on consumer devices.

Abstract

In this work, we present a novel level-of-detail (LOD) method for 3D Gaussian Splatting that enables real-time rendering of large-scale scenes on memory-constrained devices. Our approach introduces a hierarchical LOD representation that iteratively selects optimal subsets of Gaussians based on camera distance, thus largely reducing both rendering time and GPU memory usage. We construct each LOD level by applying a depth-aware 3D smoothing filter, followed by importance-based pruning and fine-tuning to maintain visual fidelity. To further reduce memory overhead, we partition the scene into spatial chunks and dynamically load only relevant Gaussians during rendering, employing an opacity-blending mechanism to avoid visual artifacts at chunk boundaries. Our method achieves state-of-the-art performance on both outdoor (Hierarchical 3DGS) and indoor (Zip-NeRF) datasets, delivering high-quality renderings with reduced latency and memory requirements.

LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering

TL;DR

This work tackles real-time rendering of large-scale 3D Gaussian Splatting on memory-constrained devices by introducing LODGE, a level-of-detail framework that builds multiple Gaussian sets and precomputes per-chunk active Gaussians to avoid per-frame recomputation. It combines depth-aware 3D smoothing and importance pruning to generate coarser levels, uses a defined active set , and partitions the scene into chunks via K-means over training camera positions; opacity blending between the two nearest chunks ensures smooth transitions. The approach also employs visibility filtering and a two-chunk blending scheme to preserve visual fidelity while dramatically reducing memory and computation. Evaluations on outdoor Hierarchical 3DGS and indoor Zip-NeRF datasets show a favorable quality-speed trade-off, enabling real-time mobile rendering with a substantially smaller memory footprint compared to prior methods. These advances make large-scale 3D reconstructions more accessible for AR/VR and navigation tasks on consumer devices.

Abstract

In this work, we present a novel level-of-detail (LOD) method for 3D Gaussian Splatting that enables real-time rendering of large-scale scenes on memory-constrained devices. Our approach introduces a hierarchical LOD representation that iteratively selects optimal subsets of Gaussians based on camera distance, thus largely reducing both rendering time and GPU memory usage. We construct each LOD level by applying a depth-aware 3D smoothing filter, followed by importance-based pruning and fine-tuning to maintain visual fidelity. To further reduce memory overhead, we partition the scene into spatial chunks and dynamically load only relevant Gaussians during rendering, employing an opacity-blending mechanism to avoid visual artifacts at chunk boundaries. Our method achieves state-of-the-art performance on both outdoor (Hierarchical 3DGS) and indoor (Zip-NeRF) datasets, delivering high-quality renderings with reduced latency and memory requirements.

Paper Structure

This paper contains 11 sections, 6 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Experiments
  5. Conclusion
  6. Supplementary Material
  7. Video
  8. Implementational Details
  9. Datasets
  10. Evaluation Protocol & Baselines
  11. Broader Impact

Figures (7)

  • Figure 1: LODGE. Applied to large-scale 3D scenes, LODGE achieves outstanding quality while maintaining superior rendering speeds. Furthermore, it enables real-time rendering on mobile devices.
  • Figure 2: Method overview.Left: The scene is represented with multiple LODs; 'active Gaussians' are selected during training based on camera distance. Right: We cluster cameras into chunks, pre-compute 'active Gaussians' per chunk, and render the two nearest chunks with 'opacity blending'.
  • Figure 3: Visible Gaussians histogram. For each pixel we compute the number of visible Gaussians and show the histogram for base model 'Full representation' and LOD rendering 'LOD'.
  • Figure 4: LOD threshold cost function. Visualized for 2 depth thresholds. Darker is lower cost.
  • Figure 5: Hierarchical 3DGS qualitative results. We compare H3DGS kerbl2024h3dgs and Octree-GS ren2024octree on SmallCity and Campus scenes. We highlight close-up region (left) and far-away region (right).
  • ...and 2 more figures